Abstract: Shearer is one of the core pieces of equipment in intelligent coal mine construction. Its manufacturing process follows a typical single-piece, small-batch production model, resulting in extremely limited data samples available for process control. In terms of component quality control, current practices primarily rely on final inspections after machining, lacking effective process early warning and real-time intervention measures. Furthermore, the complex and difficult-to-control variability caused by multiple factors, including personnel, machinery, materials, methods, and environment, further exacerbates the quality control challenges during the machining process of shearer components. This paper proposes a multistage manufacturing process control method based on digital twin. Its core lies in constructing a digital twin deeply integrated with physical mechanisms that dynamically evolves with production data. Through multidimensional mapping and simulation analysis of the manufacturing process, it enables real-time diagnosis and root-cause tracing of quality fluctuations. This method centers on a dual-drive mechanism, integrating a process knowledge model driven by physical knowledge with a data-driven graph attention network model. Within this framework, the graph network identifies critical error propagation paths and adaptively quantifies node influences, thereby proposing a weighted quality impact criticality index. By integrating multidimensional dynamic features, the method constructs an interpretable system to identify key process nodes. Its effectiveness is validated through a case study of the shearer′s rocker arm shell component. The network relationships in the manufacturing process were reduced by 19.4% after physical constraint reduction. The dynamic weight model based on Graph Attention Network (GAT) achieved an accuracy rate and F1 score of over 85% on the validation set. Key nodes were identified based on the Weighted Error Propagation Critical Index (WEPCI), and the intelligent sorting and traceability of 39 error paths were realized through the Breadth-first Search (BFS) algorithm.